Practical Paths into Applied AI

Writing as leverage

Kevin opens with writing. Brandon says he wanted to write for about ten years before he managed to stay consistent. He credits Gary Vaynerchuk’s ideas with the early push, and frames it as a systems thinker would: write once and people can read it any number of times. With an industrial and systems engineering background, he thinks in terms of bottlenecks and getting more output for less input, and a distribution channel made sense whatever the eventual goal.

Long-form writing, he says, came from his time at Amazon, where projects are argued in long written documents called the press release FAQ. He remembers writing one during his MBA internship that ran about twenty pages, single-spaced. Kevin recalls that Amazon’s executive meetings start with everyone reading the document before debating it.

Brandon admits he wants every argument to be bulletproof, which made short-form writing hard because it felt too easy to misread. AI helps here: he can rubber-duck an idea for an hour and get feedback as it grows. Today his short posts are written by hand on his phone, while newsletters often start as a 30-to-40-minute voice session on a walk, drafted with AI and then edited until the wording feels right.

Technical, non-technical, and what engineers do

For most of his career, Brandon considered himself semi-technical, because being technical meant reading and writing code and shipping to production. He now sees that line as a kind of gatekeeping. The question, in his view, is whether you are curious enough and tenacious enough to go deeper into a problem, which AI makes far easier.

He defines an engineer as someone who can keep iterating toward a solution, including when they don’t yet know how. Engineers stitch things together to make the constraints and the objective work, he says, where others stop and hand off. Kevin, noting that every earlier guest was a scientist, compares this to the scientific method: observe, form a hypothesis, change a variable, and validate the result. Brandon agrees, and adds that AI can help map the system and generate hypotheses to test.

From product manager to working with agents

As a product manager at Wonder, Brandon worked with an engineering team in Xiamen, China. Outside a weekly planning meeting, nearly everything happened in written product requirements documents and long comment threads, so he learned to anticipate what the team would be missing. A mistake cost a day. Working with agents feels similar, he says, except that it happens in real time.

He describes the skill as something like theory of mind: understanding the perspective of an agent that, like his offshore team, lacks the business context in his head. Most failures, in his view, come down to a lack of context rather than unclear intent. His contrarian habit is to ignore the advice to preserve the context window and keep talking to the agent until it can repeat back the world as he sees it, and only then start planning.

Brandon compares an agent to the lead character in the film Memento, who leaves himself tattoos and notes because he can’t form new memories. The notes are the system prompt: the agent wakes up with whatever is in front of it. He asks whether he could do the task if he woke in a dark room with a web search tool and a thin set of instructions. Kevin describes how he and James at Galang AI engineer a retrievable context layer for the same reason.

Persistent context and choosing a platform

At work, Brandon says he never starts from scratch. He begins each session by asking his agent to spin up a few sub-agents on cheap models to read his sessions from the past week, so the agent starts from what he has been doing. Kevin notes that this is also why people hesitate to commit to one personal agent: once it holds all your context, it is hard to switch. He and James keep their context layers separate from any one model, using the Hermes agent harness.

For people starting out or short on time, Brandon recommends going deep on one ecosystem instead of hopping around, since most of the learning is a mental model of how to work with these systems and manage context. He has long liked OpenAI’s ecosystem for voice, which lets far more context come through than typing, and for mobile access to Codex running on his machine. Anthropic’s models are excellent, he says, but in his view its software ecosystem is less developed for now.

Taste, convergence, and the new mean

Brandon dislikes the word taste, because it implies some people have it and others don’t. The people he sees with taste are doing the scientific method: they start with a hypothesis of what they want and iterate, with AI generating options they narrow toward the goal. That convergence, he says, is a more useful definition of good. Kevin counters that the first step, knowing and articulating what you want, is what people mean by taste. Brandon allows that taste may be what remains after you have seen enough to know what good is.

Kevin compares this to learning to write by copying great writers, or to covering songs before finding your own sound. Brandon describes using an AI agent to study short-form creators adjacent to his topics and build a guide he can try on in his own way. The risk, from his Substack post on the midpoint, is that everyone can now copy, so the new mean is still mediocre. Doing more of the same thing keeps you in the mean. Breaking out takes patience, he says, even when it looks like you aren’t doing anything.

Finding the side door

Brandon says he messaged his future manager at Vercel more than a year ago, after seeing a post about building a team of people moving between engineering and product, or who had automated work and built agents. Through circumstance, he later connected with the team. He says nearly everyone there is a better engineer than he is, but his product background lets him approach problems differently, which matters when execution is cheap.

Many job postings now draw hundreds of applicants, Brandon says, so he looks for side doors. His approach, which he has written about, works like a marketing funnel: post content to build the top of the funnel, tune your feed so you see hiring managers’ posts, and leave a short comment and a direct message when a role opens. A short comment stands out, he says, while an overeager one sends the wrong signal. A first-generation college student, he says he has always looked for angles other people weren’t seeing, and aims to raise the odds of serendipity rather than wait for it.

Tinkerers, generalists, and the cost of shipping

Kevin recalls telling his wife that coding agents are a tinkerer’s dream. His iteration loop once depended on how fast bacteria grew in an incubator; now he feels limited mainly by his imagination. Brandon says this period rewards people with exposure to many things. A product manager used to coordinate specialists and could do little alone. Now, he says, the 80th percentile is often enough because coding agents carry much of the work. Kevin describes it as breadth with access to depth.

Brandon warns that shipping has a cost. People form opinions of every tool released into an organization, and one that used to ship a few tools a year might now ship hundreds that nobody knows how to use or wants to retire. He suggests generating a hundred things and releasing two good ones. He describes an MVP that drew good feedback but still felt wrong, and another week and a half of rebuilding before it was ready.

Keeping up, and finding the bottleneck

To keep up, Brandon prunes his social feeds, marking engagement bait as not interesting and favoring people he respects who are experimenting. His mental model is a video game controller: each button does a job, and a new tool either replaces a button, if it is better, or adds a new category he needs to track.

As an industrial engineer, he also looks for the bottleneck. If something works well, there is little need to optimize it, though a jump in his primary model is worth trying right away. Kevin tells the organizations he works with that staying on today’s models for six to twelve months would still move them far ahead, and that many tasks can run on cheaper, efficient models. Brandon agrees that most organizations mainly need enablement.

Preparing students for an AI workforce

Kevin, who serves on an advisory council for his alma mater’s College of Science and Math, asks how students can prepare. Brandon suspects many assessments are broken because AI can complete them, with no replacement yet in place. He advises building skills with AI tools, developing the agency to learn whatever the problem in front of you needs, and treating getting employers’ attention as a process rather than a button to click.

Kevin argues that grades have become a weak proxy for competence and that universities need new ways to show it, along with better preparation for the job market. Brandon adds that the ability to learn new things matters more than certifying what you already know, when models and everyone else have that knowledge at hand. Kevin suggests assessments built around making something and testing it in the world, and worries about a return to Scantrons and blue books. Brandon notes how hard any new approach would be to put into practice.